Core Concepts
Gaussian Processes with contaminated normal noise improve geomagnetic perturbation predictions.
Stats
Gaussian Processes have become popular for kernel-based learning on datasets with complicated covariance structures.
A contaminated normal likelihood function is proposed to better account for heteroscedastic variance and outlier noise.
The state-of-the-art prediction model for geomagnetic ground perturbations is based on neural networks.
Sparse Variational Gaussian Process (SVGP) method is used for fitting sparse Gaussian process regression models.
Various robust Gaussian Process Regression (GPR) models with mixture noise distributions are discussed.
Inference algorithms for robust GPR models are compared.
Sparse variational GP (SVGP) method is introduced for fitting sparse GPR models with contaminated normal noise.
Simulation studies show the efficacy of the proposed inference algorithm.
Comparison of GPR models with different noise distributions on flight delays and ground magnetic perturbations data.
GPR-CN outperforms other models in handling outliers and extreme observations.
Application to flight delays dataset shows GPR-CN with the lowest NLPD and comparable RMSE and MAE.
Comparison of GPR-CN and ANN models for geomagnetic perturbation forecasting.
Quotes
"Gaussian Processes have become popular for kernel-based learning on datasets with complicated covariance structures."
"A contaminated normal likelihood function is proposed to better account for heteroscedastic variance and outlier noise."
"Sparse Variational Gaussian Process (SVGP) method is used for fitting sparse Gaussian process regression models."